Cross-Domain Multitask Learning with Latent Probit Models

Learning multiple tasks across heterogeneous domains is a challenging problem since the feature space may not be the same for different tasks. We assume the data in multiple tasks are generated from a latent common domain via sparse domain transforms and propose a latent probit model (LPM) to jointly learn the domain transforms, and the shared probit classifier in the common domain. To learn meaningful task relatedness and avoid over-fitting in classification, we introduce sparsity in the domain transforms matrices, as well as in the common classifier. We derive theoretical bounds for the estimation error of the classifier in terms of the sparsity of domain transforms. An expectation-maximization algorithm is derived for learning the LPM. The effectiveness of the approach is demonstrated on several real datasets.

On-board Robotic Multi-pinhole SPECT System for Region-of-interest (ROI) Imaging

On-board image guidance, such as cone-beam CT (CBCT) and kV/MV 2D imaging, is essential in many radiation therapy procedures, such as intensity modulated radiotherapy (IMRT) and stereotactic body radiation therapy (SBRT). These imaging techniques provide predominantly anatomical information for treatment planning and target localization. Recently, studies have shown that treatment planning based on functional and molecular information about the tumor and surrounding tissue could potentially improve the effectiveness of radiation therapy. However, current on-board imaging systems are limited in their functional and molecular imaging capability. Single Photon Emission Computed Tomography (SPECT) is a candidate to achieve on-board functional and molecular imaging. Traditional SPECT systems typically take 20 minutes or more for a scan, which is too long for on-board imaging. A robotic multi-pinhole SPECT system was proposed in this dissertation to provide shorter imaging time by using a robotic arm to maneuver the multi-pinhole SPECT system around the patient in position for radiation therapy.

A 49-pinhole collimated SPECT detector and its shielding were designed and simulated in this work using the computer-aided design (CAD) software. The trajectories of robotic arm about the patient, treatment table and gantry in the radiation therapy room and several detector assemblies such as parallel holes, single pinhole and 49 pinholes collimated detector were investigated. The rail mounted system was designed to enable a full range of detector positions and orientations to various crucial treatment sites including head and torso, while avoiding collision with linear accelerator (LINAC), patient table and patient.

An alignment method was developed in this work to calibrate the on-board robotic SPECT to the LINAC coordinate frame and to the coordinate frames of other on-board imaging systems such as CBCT. This alignment method utilizes line sources and one pinhole projection of these line sources. The model consists of multiple alignment parameters which maps line sources in 3-dimensional (3D) space to their 2-dimensional (2D) projections on the SPECT detector. Computer-simulation studies and experimental evaluations were performed as a function of number of line sources, Radon transform accuracy, finite line-source width, intrinsic camera resolution, Poisson noise and acquisition geometry. In computer-simulation studies, when there was no error in determining angles (α) and offsets (ρ) of the measured projections, the six alignment parameters (3 translational and 3 rotational) were estimated perfectly using three line sources. When angles (α) and offsets (ρ) were provided by Radon transform, the estimation accuracy was reduced. The estimation error was associated with rounding errors of Radon transform, finite line-source width, Poisson noise, number of line sources, intrinsic camera resolution and detector acquisition geometry. The estimation accuracy was significantly improved by using 4 line sources rather than 3 and also by using thinner line-source projections (obtained by better intrinsic detector resolution). With 5 line sources, median errors were 0.2 mm for the detector translations, 0.7 mm for the detector radius of rotation, and less than 0.5° for detector rotation, tilt and twist. In experimental evaluations, average errors relative to a different, independent registration technique were about 1.8 mm for detector translations, 1.1 mm for the detector radius of rotation (ROR), 0.5° and 0.4° for detector rotation and tilt, respectively, and 1.2° for detector twist.

Simulation studies were performed to investigate the improvement of imaging sensitivity and accuracy of hot sphere localization for breast imaging of patients in prone position. A 3D XCAT phantom was simulated in the prone position with nine hot spheres of 10 mm diameter added in the left breast. A no-treatment-table case and two commercial prone breast boards, 7 and 24 cm thick, were simulated. Different pinhole focal lengths were assessed for root-mean-square-error (RMSE). The pinhole focal lengths resulting in the lowest RMSE values were 12 cm, 18 cm and 21 cm for no table, thin board, and thick board, respectively. In both no table and thin board cases, all 9 hot spheres were easily visualized above background with 4-minute scans utilizing the 49-pinhole SPECT system while seven of nine hot spheres were visible with the thick board. In comparison with parallel-hole system, our 49-pinhole system shows reduction in noise and bias under these simulation cases. These results correspond to smaller radii of rotation for no-table case and thinner prone board. Similarly, localization accuracy with the 49-pinhole system was significantly better than with the parallel-hole system for both the thin and thick prone boards. Median localization errors for the 49-pinhole system with the thin board were less than 3 mm for 5 of 9 hot spheres, and less than 6 mm for the other 4 hot spheres. Median localization errors of 49-pinhole system with the thick board were less than 4 mm for 5 of 9 hot spheres, and less than 8 mm for the other 4 hot spheres.

Besides prone breast imaging, respiratory-gated region-of-interest (ROI) imaging of lung tumor was also investigated. A simulation study was conducted on the potential of multi-pinhole, region-of-interest (ROI) SPECT to alleviate noise effects associated with respiratory-gated SPECT imaging of the thorax. Two 4D XCAT digital phantoms were constructed, with either a 10 mm or 20 mm diameter tumor added in the right lung. The maximum diaphragm motion was 2 cm (for 10 mm tumor) or 4 cm (for 20 mm tumor) in superior-inferior direction and 1.2 cm in anterior-posterior direction. Projections were simulated with a 4-minute acquisition time (40 seconds per each of 6 gates) using either the ROI SPECT system (49-pinhole) or reference single and dual conventional broad cross-section, parallel-hole collimated SPECT. The SPECT images were reconstructed using OSEM with up to 6 iterations. Images were evaluated as a function of gate by profiles, noise versus bias curves, and a numerical observer performing a forced-choice localization task. Even for the 20 mm tumor, the 49-pinhole imaging ROI was found sufficient to encompass fully usual clinical ranges of diaphragm motion. Averaged over the 6 gates, noise at iteration 6 of 49-pinhole ROI imaging (10.9 µCi/ml) was approximately comparable to noise at iteration 2 of the two dual and single parallel-hole, broad cross-section systems (12.4 µCi/ml and 13.8 µCi/ml, respectively). Corresponding biases were much lower for the 49-pinhole ROI system (3.8 µCi/ml), versus 6.2 µCi/ml and 6.5 µCi/ml for the dual and single parallel-hole systems, respectively. Median localization errors averaged over 6 gates, for the 10 mm and 20 mm tumors respectively, were 1.6 mm and 0.5 mm using the ROI imaging system and 6.6 mm and 2.3 mm using the dual parallel-hole, broad cross-section system. The results demonstrate substantially improved imaging via ROI methods. One important application may be gated imaging of patients in position for radiation therapy.

A robotic SPECT imaging system was constructed utilizing a gamma camera detector (Digirad 2020tc) and a robot (KUKA KR150-L110 robot). An imaging study was performed with a phantom (PET CT PhantomTM), which includes 5 spheres of 10, 13, 17, 22 and 28 mm in diameter. The phantom was placed on a flat-top couch. SPECT projections were acquired with a parallel-hole collimator and a single-pinhole collimator both without background in the phantom, and with background at 1/10th the sphere activity concentration. The imaging trajectories of parallel-hole and pinhole collimated detectors spanned 180 degrees and 228 degrees respectively. The pinhole detector viewed a 14.7 cm-diameter common volume which encompassed the 28 mm and 22 mm spheres. The common volume for parallel-hole was a 20.8-cm-diameter cylinder which encompassed all five spheres in the phantom. The maneuverability of the robotic system was tested by navigating the detector to trace the flat-top table while avoiding collision with the table and maintaining the closest possible proximity to the common volume. For image reconstruction, detector trajectories were described by radius-of-rotation and detector rotation angle θ. These reconstruction parameters were obtained from the robot base and tool coordinates. The robotic SPECT system was able to maneuver the parallel-hole and pinhole collimated SPECT detectors in close proximity to the phantom, minimizing impact of the flat-top couch on detector to center-of-rotation (COR) distance. In no background case, all five spheres were visible in the reconstructed parallel-hole and pinhole images. In with background case, three spheres of 17, 22 and 28 mm diameter were readily observed with the parallel-hole imaging, and the targeted spheres (22 and 28 mm diameter) were readily observed in the pinhole ROI imaging.

In conclusion, the proposed on-board robotic SPECT can be aligned to LINAC/CBCT with a single pinhole projection of the line-source phantom. Alignment parameters can be estimated using one pinhole projection of line sources. This alignment method may be important for multi-pinhole SPECT, where relative pinhole alignment may vary during rotation. For single pinhole and multi-pinhole SPECT imaging onboard radiation therapy machines, the method could provide alignment of SPECT coordinates with those of CBCT and the LINAC. In simulation studies of prone breast imaging and respiratory-gated lung imaging, the 49-pinhole detector showed better tumor contrast recovery and localization in a 4-minute scan compared to parallel-hole detector. On-board SPECT could be achieved by a robot maneuvering a SPECT detector about patients in position for radiation therapy on a flat-top couch. The robot inherent coordinate frames could be an effective means to estimate detector pose for use in SPECT image reconstruction.

Bayesian Inference in Large-scale Problems

Many modern applications fall into the category of "large-scale" statistical problems, in which both the number of observations n and the number of features or parameters p may be large. Many existing methods focus on point estimation, despite the continued relevance of uncertainty quantification in the sciences, where the number of parameters to estimate often exceeds the sample size, despite huge increases in the value of n typically seen in many fields. Thus, the tendency in some areas of industry to dispense with traditional statistical analysis on the basis that "n=all" is of little relevance outside of certain narrow applications. The main result of the Big Data revolution in most fields has instead been to make computation much harder without reducing the importance of uncertainty quantification. Bayesian methods excel at uncertainty quantification, but often scale poorly relative to alternatives. This conflict between the statistical advantages of Bayesian procedures and their substantial computational disadvantages is perhaps the greatest challenge facing modern Bayesian statistics, and is the primary motivation for the work presented here.

Two general strategies for scaling Bayesian inference are considered. The first is the development of methods that lend themselves to faster computation, and the second is design and characterization of computational algorithms that scale better in n or p. In the first instance, the focus is on joint inference outside of the standard problem of multivariate continuous data that has been a major focus of previous theoretical work in this area. In the second area, we pursue strategies for improving the speed of Markov chain Monte Carlo algorithms, and characterizing their performance in large-scale settings. Throughout, the focus is on rigorous theoretical evaluation combined with empirical demonstrations of performance and concordance with the theory.

One topic we consider is modeling the joint distribution of multivariate categorical data, often summarized in a contingency table. Contingency table analysis routinely relies on log-linear models, with latent structure analysis providing a common alternative. Latent structure models lead to a reduced rank tensor factorization of the probability mass function for multivariate categorical data, while log-linear models achieve dimensionality reduction through sparsity. Little is known about the relationship between these notions of dimensionality reduction in the two paradigms. In Chapter 2, we derive several results relating the support of a log-linear model to nonnegative ranks of the associated probability tensor. Motivated by these findings, we propose a new collapsed Tucker class of tensor decompositions, which bridge existing PARAFAC and Tucker decompositions, providing a more flexible framework for parsimoniously characterizing multivariate categorical data. Taking a Bayesian approach to inference, we illustrate empirical advantages of the new decompositions.

Latent class models for the joint distribution of multivariate categorical, such as the PARAFAC decomposition, data play an important role in the analysis of population structure. In this context, the number of latent classes is interpreted as the number of genetically distinct subpopulations of an organism, an important factor in the analysis of evolutionary processes and conservation status. Existing methods focus on point estimates of the number of subpopulations, and lack robust uncertainty quantification. Moreover, whether the number of latent classes in these models is even an identified parameter is an open question. In Chapter 3, we show that when the model is properly specified, the correct number of subpopulations can be recovered almost surely. We then propose an alternative method for estimating the number of latent subpopulations that provides good quantification of uncertainty, and provide a simple procedure for verifying that the proposed method is consistent for the number of subpopulations. The performance of the model in estimating the number of subpopulations and other common population structure inference problems is assessed in simulations and a real data application.

In contingency table analysis, sparse data is frequently encountered for even modest numbers of variables, resulting in non-existence of maximum likelihood estimates. A common solution is to obtain regularized estimates of the parameters of a log-linear model. Bayesian methods provide a coherent approach to regularization, but are often computationally intensive. Conjugate priors ease computational demands, but the conjugate Diaconis--Ylvisaker priors for the parameters of log-linear models do not give rise to closed form credible regions, complicating posterior inference. In Chapter 4 we derive the optimal Gaussian approximation to the posterior for log-linear models with Diaconis--Ylvisaker priors, and provide convergence rate and finite-sample bounds for the Kullback-Leibler divergence between the exact posterior and the optimal Gaussian approximation. We demonstrate empirically in simulations and a real data application that the approximation is highly accurate, even in relatively small samples. The proposed approximation provides a computationally scalable and principled approach to regularized estimation and approximate Bayesian inference for log-linear models.

Another challenging and somewhat non-standard joint modeling problem is inference on tail dependence in stochastic processes. In applications where extreme dependence is of interest, data are almost always time-indexed. Existing methods for inference and modeling in this setting often cluster extreme events or choose window sizes with the goal of preserving temporal information. In Chapter 5, we propose an alternative paradigm for inference on tail dependence in stochastic processes with arbitrary temporal dependence structure in the extremes, based on the idea that the information on strength of tail dependence and the temporal structure in this dependence are both encoded in waiting times between exceedances of high thresholds. We construct a class of time-indexed stochastic processes with tail dependence obtained by endowing the support points in de Haan's spectral representation of max-stable processes with velocities and lifetimes. We extend Smith's model to these max-stable velocity processes and obtain the distribution of waiting times between extreme events at multiple locations. Motivated by this result, a new definition of tail dependence is proposed that is a function of the distribution of waiting times between threshold exceedances, and an inferential framework is constructed for estimating the strength of extremal dependence and quantifying uncertainty in this paradigm. The method is applied to climatological, financial, and electrophysiology data.

The remainder of this thesis focuses on posterior computation by Markov chain Monte Carlo. The Markov Chain Monte Carlo method is the dominant paradigm for posterior computation in Bayesian analysis. It has long been common to control computation time by making approximations to the Markov transition kernel. Comparatively little attention has been paid to convergence and estimation error in these approximating Markov Chains. In Chapter 6, we propose a framework for assessing when to use approximations in MCMC algorithms, and how much error in the transition kernel should be tolerated to obtain optimal estimation performance with respect to a specified loss function and computational budget. The results require only ergodicity of the exact kernel and control of the kernel approximation accuracy. The theoretical framework is applied to approximations based on random subsets of data, low-rank approximations of Gaussian processes, and a novel approximating Markov chain for discrete mixture models.

Data augmentation Gibbs samplers are arguably the most popular class of algorithm for approximately sampling from the posterior distribution for the parameters of generalized linear models. The truncated Normal and Polya-Gamma data augmentation samplers are standard examples for probit and logit links, respectively. Motivated by an important problem in quantitative advertising, in Chapter 7 we consider the application of these algorithms to modeling rare events. We show that when the sample size is large but the observed number of successes is small, these data augmentation samplers mix very slowly, with a spectral gap that converges to zero at a rate at least proportional to the reciprocal of the square root of the sample size up to a log factor. In simulation studies, moderate sample sizes result in high autocorrelations and small effective sample sizes. Similar empirical results are observed for related data augmentation samplers for multinomial logit and probit models. When applied to a real quantitative advertising dataset, the data augmentation samplers mix very poorly. Conversely, Hamiltonian Monte Carlo and a type of independence chain Metropolis algorithm show good mixing on the same dataset.